Section 12. Markov Chain Approximations

advertisement
Section 12. Markov Chain Approximations
It is often necessary to approximate models in continuous time by discrete versions. The
simplest example is approximation of Brownian motion by random walks. Let us consider
independent random variables Xi = ±1 with probability 21 each.
Sn = X1 + X2 + · · · + Xn
and
j
Sj
ξn ( ) = √
n
n
ξn is interpolated linearly between nj and j+1
n . It is not difficult to see that the distribution
of ξn (t) converges to the normal distribution with mean 0 and variance t. In fact the
joint distribution of {ξn (tj )} for any finite collection {tj } converges to the corresponding
distribution of Brownian motion at the same times, i.e. that of {β(tj )}. The proof is
nothing more than the central limit theorem. However to prove
Z ∞
x2
2
e− 2t dx
lim P [ sup ξn (s) ≥ ℓ] = P [ sup β(s) ≥ ℓ] = √
n→∞
2πt ℓ
0≤s≤t
0≤s≤t
requires some work. A general sufficient condition, which allows us to make the transition
from functions that depend on a finite number of coordinates to functionals that depend
continuously on the trajectories ( under uniform convergence) is
E[|ξn(t) − ξn (s)|4 ] ≤ C|t − s|1+α
for some α > 0 and constant C both independent of n.
More generally we can have a Markov Chain πh (x, dy) on Rd , that satisfy
Z
(yi − xi )πh (x, dy) = hbi (x) + o(h)
Z
and
(yi − xi )(yj − xj )πh (x, dy) = hai,j (x) + o(h)
Z
|y − x|4 πh (x, dy) = O(h2 )
Then the Markov Chain ξh (jh) = Xh (j) will converge to the diffusion process with generator
X
1X
Af =
ai,j (x)Dxi Dxj +
bj (x)Dxj
2
j
Let us assume that the equation
ut (t, x)+
X
1X
ai,j (x)[Dxi Dxj u](t, x)+
bj (x)[Dxj u](t, x) = 0, u(T, x) = f (x), (T = N h)
2
j
1
has a smooth solution. We define
uh (jh, x) = E[f (Xh (N h)|Xh (jh) = x] − u(jh, x)
and
∆h (jh, x) = uh (jh, x) − u(jh, x)
Then
∆h (jh, x) =
=
Z
Z
uh ((j + 1)h, y)πh (x, dy) − u(jh, x)
[uh ((j + 1)h, y) − u((j + 1)h, y)]πh(x, dy)
Z
+ [u((j + 1)h, y) − u((j + 1)h, x]πh (x, dy)
+ u((j + 1)h, x) − u(jh, x)
and
sup |∆h (jh, x)| ≤ sup |∆h ((j + 1)h, x)|
x
x
+ sup[Au((j + 1)h, x) + hut ((j + 1)h, x)]
x
+ o(h)
≤ sup |∆h ((j + 1)h, x)| + o(h)
x
Proving
sup |uh (0, x) − u(0, x)| ≤ N o(h) = o(1)
x
Examples: Biased random walk. Step size
√1 .
n
Xn+1 = Xn ±
√
√
√
1 b(j n)
π(j n, (j ± 1) n) = ± √
2
n
Converges to process with generator.
1
Dxx + b(x)Dx
2
2
√1 .
n
But
Download